Research map / Architectures and efficiency / Attention and transformers
Training–Inference Consistent Segmented Execution for Long-Context LLMs
Xianpeng Shang, Jiang Li, Zehua Duo, Qianyi Cai, Xiangdong Su
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| align training with segmented inference execution instead of full-context training (ablation) | segmented execution framework · LongBench-E | average score: 11.91 -> 24.17 | Improved |
| increase TBPTT depth from K=1 to K=2 (ablation) | segmented execution framework · LongBench-E | average score: 24.17 -> 24.07 | No clear effect |
| use training-inference consistent segmented execution vs vanilla attention | LLaMA2-7B-32K · LongBench-E summarization | score: 3.09 -> 12.10 | Improved |
| use segmented execution vs full attention on average LongBench-E | LLaMA2-7B-32K · LongBench-E | average score: 23.13 -> 23.24 | No clear effect |
| use segmented execution vs full attention on average LongBench-E | LLaMA2-7B-80K · LongBench-E | average score: 23.38 -> 24.17 | Improved |
| use segmented execution to cut prefill memory | LLaMA2-7B-80K · 32K prefill | peak memory (GB): 34.67 -> 19.06 | Improved |
| use segmented execution to cut prefill memory | LLaMA2-7B-32K · 16K prefill | peak memory (GB): 23.61 -> 18.56 | Improved |
| use segmented execution instead of vanilla attention | LLaMA2-7B-32K · 16K prefill | TTFT (s): 1.62 -> 1.70 | Hurt |
| use segmented execution on RULER beyond training length (64K) | LLaMA2-7B-32K · RULER CWE 64K | accuracy: - -> 2.00 | Improved |
| use segmented execution on RULER CWE and FWE 4K-32K | LLaMA2-7B-32K · RULER CWE/FWE avg 4K-32K | accuracy: CWE 32.94 -> 46.39; FWE 41.33 -> 43.88 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- DuoAttention — DuoAttention: Efficient Long-Context LLM Inference with Retrieval and Streaming Heads (ICLR 2025) · better on 3 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- A Little Goes a Long Way: Efficient Long Context Training and Inference with Partial Contexts (ICLR 2025 · similarity 0.91)
- CateKV: On Sequential Consistency for Long-Context LLM Inference Acceleration (ICML 2025 · similarity 0.90)
- Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation (ICLR 2025 · similarity 0.89)
- InfLLM: Training-Free Long-Context Extrapolation for LLMs with an Efficient Context Memory (NeurIPS 2024 · similarity 0.89)
- Efficient Prompt Compression with Evaluator Heads for Long-Context Transformer Inference (NeurIPS 2025 · similarity 0.89)
- A Training-Free Sub-quadratic Cost Transformer Model Serving Framework with Hierarchically Pruned Attention (ICLR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
